Seeing What Shouldn't Be There: Counterfactual GANs for Medical Image Attribution
This work addresses the limitation of existing medical image attribution methods, which focus narrowly on minimal discriminative features while neglecting other salient anatomical structures, thereby failing to provide comprehensive and clinically trustworthy explanations. To overcome this, the authors propose a class-oriented counterfactual feature attribution framework that uniquely integrates counterfactual explanation with generative adversarial networks (GANs). By incorporating a cycle-consistency loss, the method generates realistic and semantically plausible counterfactual instances that effectively highlight lesion regions critical to model decisions. The study further introduces a reliable counterfactual generation mechanism alongside dedicated quality evaluation metrics, enabling analogical self-explanation. Experiments on synthetic data, tuberculosis datasets, and the BraTS benchmark demonstrate that the generated counterfactuals exhibit superior clinical plausibility and establish a new baseline on BraTS.